When the result is not the news. Or, woe is the status of most social science reporting.
The news release was titled, “Children Of Working Moms Face More Health Problems.”
The news headlines, repeated around the world on the Internet’s instant, editor-free news rebroadcasting systems, were predictable:
Corbis helpfully sold clip art of a working mother putting her child at risk (note the ancient computer monitor and elephant-sized phone, showing the venerability of this story…):
In fact, we had the whole panoply of clip art on display, helpfully collected by Google news:
A complicated story
The research, by North Carolina State University economist Melinda Morrill, is in press at the Journal of Health Economics (the abstract is here, and Morrill has an earlier version for free on her website here).
Her conclusion, reasonably justified by the analysis was this: “I identify the effects on overnight hospitalizations, asthma episodes, and injuries/poisonings for children ages 7–17. Maternal employment increases the probability of each adverse health event by nearly 200 percent.”
Wow. Sounds awful. But understand one thing. The risk of all these events is very low, whether the kids mothers are employed or not. Doubling or tripling these rates still means that the vast majority of children are unaffected (triple-emphasis added). Using her data, a few hundred thousand National Health Interview Survey respondents from 1985 to 2004, the overall rates of each outcome look like this:
At a glance, it doesn’t match the headlines. Children whose mothers worked are less likely to be hospitalized or have asthma attacks (but more likely to have injury or poisoning). That’s probably just because healthy, rich mothers are more likely to work and have healthy, rich, safer kids. (It’s OK, we can control for that.) More importantly, the rates are low and the differences are small. Nevertheless…
The real contribution of this article is a clever application of what Belinda Luscombe at Time helpfully (I mean that) called “difficult-to-explain statistical techniques.” Specifically, Morrill used — there is no way to sugar-coat this — an instrumental variable method with two-stage least-squares regression, otherwise known as IV-2SLS.
Can this be explained to a non-expert audience? As a non-practitioner who has read a bunch of these papers, let’s see if I can do it.
- Whereas, for many questions in social science, we would like to investigate a causal relationship, but the complexity of such relationships makes that difficult to establish; and,
- Whereas, from a scientific point of view, the ideal study design is a true experiment, in which we randomly assign people to different conditions and trace their effects, thereby removing contaminating factors such as past experience, personal decisions and preferences, strengths and weaknesses, etc.; and,
- Whereas, social scientists often can’t do true experiments because of ethics (and other reasons), and when we do experiments (like laboratory simulations), they differ substantially from real-life situations;
- Therefore, some social scientists (usually called “economists”) use instrumental variable (IV) analysis, in which the trick is to find something (an “instrument”) that acts like an experiment, (more or less) randomly assigning people to different conditions, so that their true effects can be identified.
That’s the gist of it: has something (more or less) random occurred which (a) causes the independent variable to change (e.g., driving mothers into the labor force) while (b) neither causing, nor resulting from, the dependent variable (hospitalization, injury or asthma attack).
In this case, Morrill cleverly split children into two groups: those who had younger siblings who were just old enough to start kindergarten, and those who had younger siblings that were just too young to start kindergarten. Because mothers have a tendency to start work when the younger child goes off to kindergarten, but children reaching kindergarten age is neither cause nor effect of older children’s health outcomes, this acts like an experiment — some moms are assigned to the go-to-work group and some aren’t, and membership in the two groups is more or less random.
The method is called 2SLS because, using a complex prediction model, the economist first identifies those mothers whose employment was likely the result of the the younger child reaching kindergarten age, and then (in the second stage) uses a complex prediction model to determine whether those mothers’ older children were more likely to end up sick or injured than the children of those who did not start work.
This method creates something close to an experiment, close enough that it is sometimes called a “natural experiment,” since the scientist didn’t engineer it. However, it also analyzes events that are extremely narrowly construed. It really is only a test of what happens when American mothers of two or more children started work after the younger child reached kindergarten age (over a 20-year period) — holding constant a wide array of social and demographic variables. Since the randomness of the school-age “instrument” can, in a practical sense, be confirmed statistically, the effect is reasonably called “causal,” but caution in the interpretation is wise.
And Morrill was cautious. Although she found that older children of mothers who went off to work in these conditions were indeed more likely to suffer these ill effects, she did a number of other checks to make sure things were as they appeared.
In fact, one of the mostly-overlooked aspects of the paper was a section on “heterogeneous effects.” Here, she tested whether the overall effect she found actually resulted from some subset of the families experiencing large effects while others experienced none. In my interpretation, this is where the real story is.
The effect of mothers going off to work on children’s hospitalization was three-times greater for Blacks than for Whites (and non-existent for Hispanics). The effect was only significant for mothers who had no more than high school education (unlike most or all of the women in the clip art above!). And the effect was three-times larger for single mothers than married mothers.
With no measures of child care availability or any details about the care arrangements of the families’ children, I’m left to conclude that the results probably reflect the simple fact that poorer women have fewer good options for childcare, so that when they enter the labor force, their children experience some increased risk of accident or illness.
Stop the presses
I see this result as a confirmation of common sense, not shocking or disturbing, or in any substantive way altering my understanding of the work-family-children situation: mothers working for pay increases the risks of illness or injury associated with non-supervision, or supervision by others. That this seems obvious does not detract from the value of the study, just from the breathlessness of its news coverage.
What are the implications of this? I can think of two. First, mothers (or, obviously, any caretakers) who are considering entering the labor force need to consider the availability and quality of alternative care arrangements for the children they will no longer be caring for during their working hours. Hopefully, they already knew this. Second, for public policy, we need to consider the availability and affordability of care arrangements for children whose parents are employed.
As for the bigger question, the one about mothers’ guilt and hard choices, Belinda Luscomb was good enough to link to a recent meta-analysis — a study of studies — that analyzed 69 different studies of the effect of mothers’ early employment on their children’s school achievement and psychological health, published by the American Psychological Association. That study concluded:
The small effect size and primarily nonsignificant results for main effects of early maternal employment should allay concerns about mothers working when children are young. However, negative findings associated with employment during the child’s first year are compatible with calls for more generous maternal leave policies. Results highlight the importance of social context for identifying under which conditions and for which subgroups early maternal employment is associated with positive or negative child outcomes.
Now we can confirm that another risk — small and manageable in the vast majority of cases — is illness or injury associated with loss of parental supervision. Something to watch out for. But didn’t we already know that?
There, I said it. Sorry it took so long.